Single-cell technologies have transformed molecular biology by enabling high-resolution profiling of cellular heterogeneity across tissues. However, the growing scale and complexity of single-cell datasets introduce significant analytical challenges, including batch effects across experiments, limited resolution in marker gene selection, and the high cost of spatial transcriptomics. This thesis addresses these challenges through three complementary methodological innovations. First, we develop a supervised integration framework using domain adaptation to align scRNA-seq datasets across batches while preserving rare and batch-specific cell types. By incorporating cell-type annotations into a neural embedding model, the approach ensures more accurate and biologically meaningful alignment across studies. Second, we propose a hierarchical marker gene selection strategy that moves beyond the limitations of one-vs-all differential expression. By organizing cell types into lineage-informed hierarchies, the method identifies markers at varying levels of granularity, improving resolution and interpretability in complex cell populations. Third, to address the cost and scalability barriers of spatial transcriptomics, we introduce a deep learning pipeline that predicts cell-type composition directly from H&E-stained histology images. Using embeddings from pre-trained foundation models, the approach trains a lightweight MLP regressor to estimate cell-type proportions without relying on gene expression measurements, enabling scalable spatial analysis at reduced cost. Together, these contributions offer a unified framework for enhancing integration, annotation, and spatial inference in single-cell studies, improving accessibility and analytical power across diverse experimental settings.